openreview-explorer

name	openreview-explorer
description	Search papers on OpenReview and retrieve peer reviews and rebuttals from major ML conferences (ICLR, NeurIPS, ICML). Core skill for FrontierPilot's Knowledge Exploration engine.
version	1.0.0

OpenReview Explorer

This skill fetches papers, peer reviews, and author rebuttals from OpenReview — treating the review process as structured learning material for newcomers entering a research field.

Why Reviews Matter

Peer reviews reveal what the field actually cares about: common weaknesses, evaluation standards, and how top authors respond to criticism. This is knowledge that takes years to accumulate informally — FrontierPilot surfaces it directly.

Python Interpreter

Always use: python3 (available in the OpenClaw container environment)

Capabilities

1. Search papers by keyword + venue

python3 /home/node/.openclaw/workspace/skills/openreview-explorer/scripts/search_papers.py \
  --query "automl" \
  --venue "ICLR" \
  --year 2024 \
  --limit 5

2. Get reviews + rebuttal for a paper

python3 /home/node/.openclaw/workspace/skills/openreview-explorer/scripts/get_reviews.py \
  --forum-id <FORUM_ID> \
  --venue "ICLR" \
  --year 2024

3. Get conference papers (all from a venue/year)

python3 /home/node/.openclaw/workspace/skills/openreview-explorer/scripts/search_papers.py \
  --venue "ICLR" \
  --year 2024 \
  --limit 10

Workflow for Knowledge Exploration

1. User provides a topic (e.g., "diffusion models for image generation")
2. Run search_papers.py to find top relevant papers from recent ICLR/NeurIPS/ICML
3. For each promising paper, run get_reviews.py to retrieve:
   • Reviewer ratings and summaries
   • Detailed strengths/weaknesses
   • Author rebuttal (how the authors defended their work)
4. Synthesize: What do reviewers consistently praise? What weaknesses appear? What does the rebuttal reveal about the authors' intent?
5. Save synthesis to memory/RESEARCH_LOG.md

Output Format

Papers are returned as JSON. Present them as:

• Title, authors, venue, year
• Abstract excerpt
• Forum ID (needed for get_reviews.py)

Reviews are returned with:

• Rating (e.g., "8: accept, good paper")
• Summary, strengths, weaknesses
• Confidence score
• Rebuttal excerpt

Environment Variables

The scripts read credentials from environment variables. Set them before use:

• OPENREVIEW_USERNAME: your OpenReview account email
• OPENREVIEW_PASSWORD: your OpenReview account password
• OPENREVIEW_BASE_URL: https://api2.openreview.net (default, usually no need to set)

Supported Venues

Venue	Invitation Pattern
ICLR 2024	ICLR.cc/2024/Conference/-/Submission
ICLR 2023	ICLR.cc/2023/Conference/-/Submission
NeurIPS 2024	NeurIPS.cc/2024/Conference/-/Submission
ICML 2024	ICML.cc/2024/Conference/-/Submission

Error Handling

• If a paper has no reviews yet (e.g., under review), the script returns an empty list — note this to the user
• Rate limit: space out requests for large batches (>20 papers)
• If login fails, fall back to anonymous mode (public papers only, no reviews)